A Simplified Conjoint Recognition Paradigm for the Measurement of Gist and Verbatim Memory

Transcription

1 Journal of Experimental Psychology: Learning, Memory, and Cognition 2008, Vol. 34, No. 3, Copyright 2008 by the American Psychological Association /08/$12.00 DOI: / A Simplified Conjoint Recognition Paradigm for the Measurement of Gist and Verbatim Memory Christoph Stahl and Karl Christoph Klauer University of Freiburg The distinction between verbatim and gist memory traces has furthered the understanding of numerous phenomena in various fields, such as false memory research, research on reasoning and decision making, and cognitive development. To measure verbatim and gist memory empirically, an experimental paradigm and multinomial measurement model has been proposed but rarely applied. In the present article, a simplified conjoint recognition paradigm and multinomial model is introduced and validated as a measurement tool for the separate assessment of verbatim and gist memory processes. A Bayesian metacognitive framework is applied to validate guessing processes. Extensions of the model toward incorporating the processes of phantom recollection and erroneous recollection rejection are discussed. Keywords: memory, conjoint recognition, multinomial modeling, source monitoring, metacognitive guessing strategies In research on episodic memory, fuzzy trace theory (FTT) has recently received much attention. At its core lies the assumption of separate memory traces for the meaning or gist of an item (e.g., its semantic category) and its identity or verbatim detail (e.g., its exact wording) that can be retrieved independently from memory. This distinction has proven fruitful in a number of domains but especially so in research on false memory (e.g., Brainerd, Forrest, Karibian, & Reyna, 2006; Brainerd, Payne, Wright, & Reyna, 2003; Brainerd & Reyna, 2002; Brainerd, Wright, Reyna, & Mojardin, 2001; Odegard & Lampinen, 2005; Seamon et al., 2002; Wright & Loftus, 1998). To empirically separate verbatim and gist memory, Brainerd, Reyna, and Mojardin (1999) have proposed the conjoint recognition (CR) paradigm and multinomial model. However, only few studies have since used this relatively complex paradigm. In the present research, a simplified CR paradigm is introduced and validated. In Brainerd et al. s (1999) CR paradigm, participants are first presented with a study list. The test list contains three types of items: target probes (i.e., old items from the study list), related distracters that share a target s gist, and unrelated distracters. The memory test is administered to three groups of participants with different instructions: Under the T instruction, participants are asked to accept as old only targets; under the R instruction, Christoph Stahl and Karl Christoph Klauer, Institute for Psychology, University of Freiburg, Freiburg, Germany. The research reported in this article was supported by Deutsche Forschungsgemeinschaft Grant Kl 614/31-1 to Karl Christoph Klauer. We would like to thank Stephan Stegt for providing the materials used in Experiment 2, and William Batchelder, Charles Brainerd, John Dunn, Stephan Lewandowski, and Timothy Odegard for their valuable comments on drafts of this article. Correspondence concerning this article should be addressed to Christoph Stahl, Institute for Psychology, University of Freiburg, D Freiburg im, Breisgau, Germany. stahl@psychologie.unifreiburg.de participants are to accept as old only related distracters; and, finally, under the T R instruction, participants are to accept both targets and related distracters. From the proportions of accepted targets, related distracters, and unrelated distracters obtained in the three different groups, the parameters of a model are estimated that provide measures of verbatim and gist memory. In the following, a simplified CR paradigm and model are proposed that also provide valid estimates of verbatim and gist memory but on the basis of a much simpler procedure. In the simplified CR paradigm, all participants are presented with the same test list, and a single instruction is presented asking the same multiple choice question for all items. As we will show, in the simplified CR paradigm, valid measures of gist and verbatim memory can be obtained from a single group of participants. The Original CR Paradigm and Model In the original version of the CR paradigm, three different memory processes are empirically separated by way of a multinomial measurement model: an identity judgment based on the verbatim trace, a similarity judgment based on the gist trace, and a process of recollection rejection (Brainerd, Reyna, Wright, & Mojardin, 2003). FTT postulates that for each event, two different memory traces are created: First, a verbatim trace stores the event s perceptual detail. At test, when a match is found between a verbatim trace and the verbatim information present in the probe, an identity judgment is made that leads to the probe s acceptance. Second, independently, the gist trace stores the core meaning of an item. At test, a detection of similarity between the probe s gist and the gist traces stored in memory is assumed to lead to an acceptance response. Both memory traces are stored separately and can be retrieved independently. According to FTT, false memories arise when the gist trace of an event is not integrated with its verbatim trace, that is, when gist memory is retrieved but verbatim memory is not. In this case, two opposing processes are postulated. First, a gist-based similarity judgment is thought to underlie false recognition of related dis- 570

2 A SIMPLIFIED CONJOINT RECOGNITION PARADIGM 571 tracters. This process is in opposition to a second verbatim-based recollection rejection process that is the basis of rejecting related distracters. Recollection rejection is thought to be based on judging nonidentity between the verbatim information present in the probed related distracter and the verbatim trace of the corresponding target, and thereby reduces false recognition of semantically related lures based on item-specific verbatim traces (Brainerd et al., 1999; Brainerd, Reyna, et al., 2003). Considerable evidence for this process has accrued (e.g., Lampinen, Odegard, & Neuschatz, 2004; Lampinen, Watkins, & Odegard, 2006; Seamon et al., 2002). To separate verbatim and gist memory traces empirically, Brainerd and colleagues have proposed the CR paradigm (Brainerd, Holliday, & Reyna, 2004; Brainerd et al., 1999, 2001; Brainerd, Stein, & Reyna, 1998). In a CR memory test, participants are presented with targets (i.e., items from the study list), related distracters (i.e., items that had not been presented on the study list but are related to a target via common gist), and unrelated distracters (i.e., new items that were neither part of the study list nor are related to a target). In the original CR paradigm introduced by Brainerd et al. (1999), participants undergo a memory test under one of three instruction conditions: the T, R, or T R conditions that have already been introduced above. From the 3 (probe types) 3 (instruction conditions) acceptance probabilities, the parameters of a multinomial model are estimated that provide measures of gist and verbatim memory as well as estimates of acceptance by guessing (for an introduction to multinomial models, see Batchelder & Riefer, 1999; Riefer & Batchelder, 1988). Performance in the CR memory test is determined by the interplay of the identity, similarity, and recollection rejection judgment processes, as well as by guessing. The processing-tree representation of the original multinomial CR model (see Figure 1) illustrates how the postulated processes interact. Consider the first tree diagram that represents cognitive processes in reaction to a target probe. When the verbatim trace of the target can be retrieved (with probability V t ), comparison with the probe yields an identity judgment, which can be conceived of as a conscious and explicit recollection of the target episode. Participants can therefore correctly accept a target under the T and T R instructions and correctly reject it under the R instruction. With probability 1 V t, the verbatim trace cannot be retrieved. In that case, the gist trace can still be retrieved with probability G t, resulting in a judgment of similarity between the gist of the probe and the gist of a memory episode. Given this mnemonic state the meaning of an item has successfully been retrieved, but no information is available that would support a classification of that item as a target or a related probe which response should be selected? Brainerd et al. (1999) postulated that participants automatically attribute the detected similarity to the source specified in the instruction (i.e., under the T instruction, a similarity judgment would lead to a target response, whereas under the R instruction, it would result in a related response). Brainerd et al. (1999) reported indirect evidence to support this assumption. Thus, a similarity judgment is assumed to produce an acceptance response under all three instruction conditions. With probability (1 V t )(1 G t ), neither the verbatim nor the gist trace can be retrieved, and participants guess whether to accept (with a different probability b i for each instruction condition) or reject the probe (with probability 1 b i ). Next, consider the second tree diagram in Figure 1 for a related probe. First, it is possible that the related probe acts as a retrieval cue for the verbatim trace of the target that it is related to (with probability V r ). In this case, a judgment of nonidentity between the verbatim information of probe and target results in the rejection of the probe under the T instruction and acceptance under the R and Figure 1. Processing tree model for the original conjoint recognition paradigm. Rectangles on the left denote probe type, rectangles on the right denote responses; columns represent different instruction conditions. Branches of the processing tree represent the combination of cognitive processes postulated by the model. V t probability of retrieving a target s verbatim trace given a target probe; V r probability of retrieving a target s verbatim trace given a related probe; G t probability of retrieving a target s gist trace given a target probe; G r probability of retrieving a target s gist trace given a related probe; b probability of guessing that an item is old. Note that different parameters b T, b R, and b T R are used to represent guessing under the different instruction conditions (i.e., under the T instruction, participants are asked to accept as old only targets; under the R instruction, participants are to accept as old only related distracters; and under the T R instruction, participants are to accept both targets and related distracters).

3 572 STAHL AND KLAUER T R instructions. With probability 1 V r, the corresponding target s verbatim trace cannot be retrieved. In that case, its gist trace can still be retrieved with probability G r, which would provoke a similarity judgment causing acceptance under all conditions. With probability (1 V r )(1 G r ), neither the corresponding target s verbatim trace nor its gist trace can be retrieved, and responses are again determined by guessing processes. As illustrated by the third diagram in Figure 1, when presented with an unrelated probe, neither verbatim nor gist traces are available, and participants responses are assumed to be based solely on guessing processes. This implies that the original CR model does not incorporate a process by which new items can be detected as new (e.g., a metacognitive process, such as the one suggested by Strack & Bless, 1994). Models of recognition memory differ with regard to the inclusion of such a process (for discussions of this issue for source monitoring [SM] models, see, e.g., Batchelder & Riefer, 1990; Batchelder, Riefer, & Hu, 1994; Bayen, Murnane, & Erdfelder, 1996). The model s parameters can be interpreted as the probability of the cognitive processes that they represent. Thus, V t represents the probability of retrieving a target s verbatim trace, given the target probe as a retrieval cue. V r represents the probability of retrieving a target s verbatim trace, given a related probe as a retrieval cue. V r is assumed to be smaller than V t because a related probe likely does not constitute as good a retrieval cue as the target itself. G t and G r represent the probabilities that responses to a target probe and a related probe, respectively, are based on similarity ratings in the absence of retrieval of a verbatim trace. For these parameters, no order relation has been proposed a priori. Finally, b T, b R, and b T R represent the probabilities of accepting a probe via guessing processes in the T, R, and T R conditions, respectively. In a nutshell, in the original CR paradigm, participants recognition memory for three types of probes (targets, related distracters, and unrelated distracters) is tested under three betweensubjects conditions. A multinomial model is then fitted to the data, and its parameters provide estimates of verbatim and gist memory traces as well as guessing processes. We endorse the model-based approach to process separation that was followed in the original CR paradigm. However, we believe that the same result separate and uncontaminated measures of verbatim and gist memory as well as guessing processes can be achieved more efficiently. Here, we propose a simplified CR paradigm in which a single memory test condition suffices to obtain those estimates. 1 The Simplified CR Paradigm and Model The CR procedure can be modified so that it is no longer necessary to administer a memory test in three separate groups, implying a considerable advantage in terms of the efficiency of data collection. In the present simplified CR paradigm, estimates for verbatim and gist memory can be obtained for a single group of participants, as compared with three groups of participants in the original paradigm, thereby reducing the costs in terms of the required number of subjects by two thirds. In the simplified paradigm, as in the original paradigm, participants are presented with targets, related probes, and unrelated probes, and they are informed as to the types of items that compose the test list. However, in contrast to the original paradigm, the simplified paradigm asks not for acceptance or rejection responses but for the identification of the probe s type. Participants are asked to classify targets, related probes, and unrelated probes in a single condition. They are instructed to respond target if they believe that the current probe has been presented in the learning phase. If they believe the current probe to be a related distracter, they are instructed to indicate this by selecting the related response. If they consider the probe to be an unrelated distracter, they are to select the new response. This procedure results in 3 2 independent empirical probabilities (two response probabilities for each type of probe are free to vary). Before we introduce the multinomial model for the simplified paradigm, note that the procedural simplification affects the processes that are thought to occur given the mnemonic state of gist memory without verbatim memory. Given gist memory without verbatim memory, participants infer that the item is not an unrelated probe, but they are left to choose between the target and related responses. As mentioned above, in this case, it was assumed that the similarity judgment would automatically produce an accept response in all three conditions of the original CR paradigm (Brainerd et al., 1999). This assumption is no longer necessary in the simplified procedure. Instead, a new guessing parameter a is introduced to model the probability with which participants select the response target rather than related in this mnemonic state. The modified multinomial model for the simplified CR paradigm has a total of six parameters. Given six empirical probabilities, it is a saturated model. The equations and a proof of identifiability are provided in the Appendix, in which we also discuss the relationship of the CR model to the related SM model. The processing-tree representation of the model is given in Figure 2. It is almost identical to the original model, with the exception of an additional parameter a that represents the process of guessing target or related. This process is relevant in case of available gist but no verbatim memory, (1 V t )G t, and in the case that a probe s meaning has been classified as old by way of guessing, (1 V t )(1 G t )b. Consider the first tree diagram that represents the cognitive processes occurring when a target probe is presented at test. In case of available verbatim memory, it is correctly identified as a target. Given no 1 Despite its ability to separately assess verbatim and gist processes, the original CR paradigm and model has rarely been implemented (a database search for the term conjoint recognition yielded only 12 hits, as opposed to 129 hits for fuzzy trace theory; for exceptions, see Brown & Gorfein, 2004; Rotello, 2001; Stahl, 2004, 2006; Stahl & Klauer, in press). This might be due to the relatively costly design, requiring three between-participants test conditions for each single level of the independent variable. In addition, the relative novelty of the multinomial modeling approach and the potentially problematic assumption of the identity of cognitive processes across conditions (e.g., Rotello, 2001) might be relevant. Cowan (1998) suggested that the R instruction may be too complex and might therefore not yield reliable results. In line with this claim, in our own research using the CR paradigm (Stahl, 2004, 2006), we regularly observed that the R instruction posed a challenge for a (small) number of individuals from our sample of college students. If the R instructions proved to be too complex, data from the original CR procedure are likely to be unreliable. Note, however, that the CR procedure has been successfully applied not only for adults but also for children as young as 5 years of age (e.g., Brainerd et al., 1998, 2004), suggesting that, whereas specific realizations of it might be problematic, this is not true for the R instruction per se.

4 A SIMPLIFIED CONJOINT RECOGNITION PARADIGM 573 Figure 2. Processing tree model for the simplified conjoint recognition paradigm. Rectangles on the left denote probe type, rectangles on the right denote responses. They are connected by branches of the processing tree that represent the combination of cognitive processes postulated by the model. V t probability of retrieving a target s verbatim trace given a target probe; V r probability of retrieving a target s verbatim trace given a related probe; G t probability of retrieving a target s gist trace given a target probe; G r probability of retrieving a target s gist trace given a related probe; b probability of guessing that an item is either a target or a related probe; a probability of guessing target. verbatim memory but available gist memory participants have identified the probe s meaning as old but cannot remember whether the probe itself or a related item with the same gist had been presented in the learning phase a decision has to be made between the target and related response options. With probability a, the probe is classified as a target, and with probability 1 a, the probe is classified as a related distracter. Should neither memory trace be available, participants can still guess that the probe s meaning is old (with probability b). In this case, a choice between the target and related responses is again required, which is again modeled by the new parameter a as described above. The same decision process involving parameter a is postulated to occur for related probes, as can be seen in the branches (1 V r )G r and (1 V r )(1 G r )b of the second diagram. Classifications of unrelated probes are based on a combination of guessing processes a and b, as illustrated in the third diagram. Note that, as in the original CR model, we do not include a process of detecting new items as new in the simplified model. However, it is possible to do so, and we discuss this possibility in the General Discussion section. Validation of a Measurement Model A few words are in order to clarify the status of the models. In our view, the mathematical models discussed here should not be taken as precise and complete theories of recognition memory. In line with Batchelder and Batchelder (2008), we view models as measurement tools that are useful in empirically dissociating cognitive processes of interest. Such measurement tools are useful to the extent that they provide good approximations of the cognitive processes of interest. It is therefore necessary to demonstrate for each such model that its parameters are valid measures of the processes they represent. This can be achieved in an empirical validation program in which each of the processes represented in the model are separately targeted by manipulations that are hypothesized to selectively affect the given process. A model can be considered valid when its parameters are shown to selectively respond to manipulations targeted at the process that is measured by that specific parameter. In the following, we will present such a validation program for the model of the simplified CR paradigm to demonstrate that the model for the simplified CR paradigm is capable of separating verbatim and gist memory. We conducted a series of six studies using the simplified CR paradigm. In these studies, experimental manipulations are implemented that have been shown to affect verbatim and gist memory parameters in the original CR paradigm by Brainerd et al. (1999), and results demonstrate that these manipulations have the same effects on the parameters for verbatim and gist memory obtained through the simplified CR paradigm. Participants General Method Participants were sampled from the department s database of volunteers (mostly students from Freiburg s universities and colleges, as well as nonstudent citizens) and participated in exchange for a certificate of participation or monetary compensation (Ex-

5 574 STAHL AND KLAUER periments 1 and 2: 7 Euro; Experiments 3 6: 3.50 Euro). Each volunteer participated in only one of the reported experiments. Participants native language was German. Materials Three sets of German word lists were used: synonym pairs, category lists, and Deese Roediger McDermott (DRM; Deese, 1959; Roediger & McDermott, 1995) lists. In all experiments, five items were added as primacy buffer, and five items were added as recency buffer to the study lists. Synonym pairs. One set consisted of synonym word pairs with each pair denoting one occupation concept. Gist memory was defined as memory for the concept, and verbatim memory was defined as memory for the word that was presented. Twenty pairs of occupation names were taken from Stahl (2004), in which both members of a pair were synonyms for a single target occupation (e.g., barber and hairdresser). Pairs were randomly assigned to contribute either a target or a related probe to the test list, each synonym of a given pair being selected as a target or related probe with the same probability. Twenty additional occupation names were used as unrelated probes that denoted other occupations. Category exemplars. The second set consisted of exemplar pairs of common categories. Gist memory was defined as memory for the category, and verbatim memory was defined as memory for the presented exemplar. Two exemplars were generated for 30 common categories by the authors (e.g., hammer and saw as exemplars of the category tools). Categories were randomly assigned to contribute either a target, a related probe, or an unrelated probe to the test list, and each exemplar was selected as a target, related, or unrelated probe with the same probability. DRM lists. A third set was used in Experiment 2 and consisted of German DRM lists (Deese, 1959; Roediger & McDermott, 1995) that were taken from Stegt (2006). DRM lists consist of a number of to-be-presented items and a single critical lure item that is related to list items by common gist. The critical lure items are not presented but nevertheless often recalled and recognized with high probability. Gist memory was defined as memory for the list theme, and verbatim memory was defined as memory for the presented list item. Thirty DRM lists with four list items and one critical lure each were used. Lists were randomly assigned to contribute either a target, a related distracter, or an unrelated distracter to the test list. The first list items of presented lists were used as targets, the critical lures were used as related distracters, and the first list items of nonpresented lists were used as unrelated distracters. Experiments 1 and 2: Gist Memory In Experiments 1 and 2, we aimed to validate gist memory parameters. We implemented a manipulation that has been shown to affect gist memory in the original CR paradigm. Brainerd et al. (2001, Experiment 3) manipulated gist memory for a concept by presenting one versus multiple items related to that concept at study. When the concept was repeatedly activated at study by multiple items, memory for its gist was increased. In Experiment 1, gist memory was manipulated by varying the number of items (one vs. four) that were presented at study from a given target category. Gist memory was predicted to be greater for targets as well as for related items from categories from which more exemplars were presented. In Experiment 2, we attempted to selectively target the gist memory parameter for related items, as this process is thought to underlie many phenomena of false memory (e.g., Brainerd & Reyna, 2002). We again presented one versus four items from a given concept, but in this study, we used DRM lists (Deese, 1959; Roediger & McDermott, 1995) that are typically used in false memory research. DRM lists consist of semantic associates converging on critical, nonstudied lure words (e.g., butter, loaf, knife bread). In this paradigm, false alarms for nonpresented, critical lure words (e.g., bread) are often as frequent or even more frequent than hits for presented items (e.g., butter). An important characteristic of these lists is that critical lures are strong semantic associates of list items, whereas strong semantic associations do not necessarily exist among list items. As a result, by presenting an increasing number of list items, the critical lure s gist becomes increasingly activated, whereas the gist of other list items is not or only weakly affected. We therefore predicted increased gist memory parameters for related probes (i.e., the critical lures) but not for target probes (i.e., the first item on each list). Method Participants. Twenty volunteers participated in Experiment 1. One participant had to be excluded because of prior participation in another experiment of the series. Nineteen participants remained (13 women; ages ranged from 20 to 39 years, M 23). In Experiment 2, 21 volunteers participated (13 women; ages ranged from 18 to 23 years, M 20). Design. A 2 (Gist Activation: weak vs. strong) 3 (Probe Type: target, related, unrelated) repeated-measures design was implemented. Materials. Gist Activation was manipulated by presenting different numbers of items from each concept on the study list. In the weak condition, a concept was represented by a single item; in the strong condition, a concept was represented by four items. Experiment 1 used the 30 categories described in the General Method section. Two additional category exemplars were generated for each category so that there were five exemplars per category. The study list presented exemplars from 20 randomly selected categories that were randomly split into two halves of 10 categories each. Categories from the first half were represented by a single item randomly drawn from the five available exemplars; categories from the second half were represented by four items, again randomly drawn from the five available exemplars. In total, 50 items were thereby presented in random order. At test, the 10 single-item categories were randomly split into five categories for which the single presented item was shown as target probe and five categories for which a randomly selected nonpresented exemplar was shown as related probe. Similarly, the 10 four-item categories were randomly split into five categories for which a randomly selected item from the four presented exemplars was shown as the target probe and five categories for which the nonpresented fifth exemplar was shown as the related probe. Ten unrelated probes were randomly selected from the exemplars of the remaining categories that were not presented in the study list, one unrelated probe representing each such category. Order of presentation of these 30 test list items was randomized. All randomizations were carried out for each participant anew.

6 A SIMPLIFIED CONJOINT RECOGNITION PARADIGM 575 In Experiment 2, we used 30 DRM lists (see the General Method section). The study list presented items from 20 randomly selected DRM lists that were randomly split into two halves of 10 DRM lists each. DRM lists from the first halves were represented by the first DRM list item; DRM lists from the second half were represented by the first four items. In total, 50 items were thereby presented in random order. At test, the 10 single-item DRM lists and the 10 four-item DRM lists were each randomly split into five DRM lists for which the first DRM list item was shown as target probe and five DRM lists for which the critical lure was shown as related probe. Unrelated probes were the first DRM list items from 10 DRM lists that were not represented in the study list. In total, 30 items were thereby presented at test in random order. All randomizations were carried out for each participant anew. Procedure. In Experiments 1 and 2, we used the simplified CR procedure in individual computerized sessions. Participants were instructed that they were to be presented with a list of items that they were to remember for a later test. Study items were presented sequentially for 4,000 ms in black Sans-Serif letters on a gray background in the center of the screen. After the study phase, participants solved arithmetic problems for a total duration of 5 min. The memory test was administered approximately 24 hr after the learning phase to minimize effects of verbatim memory. In the memory test, participants were presented sequentially with a list of probes, and they indicated their mnemonic state for each probe by selecting the appropriate response with a computer mouse. Specifically, they were to indicate whether the probe was identical to an old item (i.e., a target ), related to an old item, or new. On the basis of the results from two pilot studies, we chose to present participants with a simultaneous decision with these three response options (see also Marsh & Hicks, 1998). After completing the memory test, participants were thanked, debriefed, and dismissed. Results Response frequencies are given in Table 1. Parameter estimates and significance tests are given in Table 2. Guessing parameters were set equal across the Gist Activation factor for identifiability reasons. 2 Parameter estimation and hypotheses tests reported below were performed with the HMMTree software (Stahl & Klauer, 2007). Sensitivity power analyses (performed with G*Power 3; Faul, Erdfelder, Lang, & Buchner, 2007) assured high test power, 1.95, for parameter comparisons across conditions. With.05, we were able to detect small to medium effects (.09 w.25; see Cohen, 1988, chapter 5). In Experiment 1, as predicted, gist memory (parameters G t and G r ) was greater for items from categories from which four exemplars had been presented at study than for items from categories from which only a single item had been presented. In Experiment 2, again as predicted, gist memory for related probes (G r ) but not for targets (G t ) was greater for items from DRM lists from with four items had been presented at study than for items from DRM lists from which only a single item had been presented. No other effects were significant. Verbatim memory was not affected. Although estimates of verbatim memory (V t ) are slightly increased for targets from four-item categories and DRM lists, this difference is not significant. Estimates of the recollection rejection process (V r ) appeared to be reduced for four-item categories in Table 1 Observed Frequencies of Memory Judgments in Experiments 1 6 Experiment Manipulation Probe type Experiment 1, but this is likely due to random error, given the large confidence intervals of V r parameters. Discussion Response t r u Experiment 1 Gist Activation: Weak t r Gist Activation: Strong t r u Experiment 2 Gist Activation: Weak t r Gist Activation: Strong t r u Experiment 3 Target Presentations: 1 t r Target Presentations: 2 t r u Experiment 4 Target Presentations: 1 t r Target Presentations: 2 t r u Experiment 5 Target-first t r u Target-last t r u Control t r u Experiment 6 Target-first t r u Target-last t r u Control t r u Note. t target probe; r related distracter; u unrelated distracter. In Experiments 1 and 2, we examined the effects of a manipulation known to increase gist memory on the parameters of the CR model. In Experiment 1, activation of the gist of a category was 2 It was necessary to introduce additional restrictions because the manipulation of number of presentations affected only targets and related probes but did not affect unrelated probes. Therefore, it yielded only four additional independent empirical probabilities, and as a result, it was possible to estimate only four additional parameters, making it necessary to equate two parameters across conditions. This is in contrast to the betweenparticipants manipulations used in Experiments 5 and 6 that yielded six additional empirical probabilities for each condition.

7 576 STAHL AND KLAUER Table 2 Estimates (and 95% Confidence Intervals) for the Parameters of the Simplified Conjoint Recognition Model for Experiments 1 and 2 Parameter Weak Gist activation Strong Experiment 1 a.25 (.11,.40) b.17 (.12,.23) G t.33 (.10,.56).76 (.56,.96) G r (.07,.63) (.76,.98) V t (.51,.73) V r.20 (.00,.51) (.62,.85).00 (.00,.62) Experiment 2 a.41 (.30,.52) b.37 (.30,.43) G t.25 (.00,.50).54 (.30,.77) G r (.00,.47) (.65,.90) V t (.28,.56) V r.00 (.00,.22) (.40,.69).00 (.00,.31) G 2 (df 1) manipulated by varying the number of items (one vs. four) that was presented from each category at study. Results indicate that this manipulation affected the gist parameters as expected: Both the gist parameters for targets and for related items were increased for categories from which four items had been presented. In Experiment 2, we aimed to discriminate between gist memory for targets and related probes by asymmetrically affecting the different types of gist memory. To accomplish this, we used DRM lists (Deese, 1959; Roediger & McDermott, 1995). These lists are characterized by strong associations from list items to a critical item, whereas associations between list items are weaker. To manipulate gist memory, we presented participants at study with either one or four DRM list items. At test, we presented the critical item as related probe, and we expected that gist memory for critical items (i.e., G r ) should be increased strongly by the presentation of three additional associates. On the other hand, gist memory should only show a modest increase for the first item from each list that was presented as a target probe (i.e., G t ). This pattern was observed in gist parameters for Experiment 2. In sum, gist memory parameters of p Note. a probability of guessing target ; b probability of guessing that an item is either a target or a related probe; G t probability of retrieving a target s gist trace given a target probe; G r probability of retrieving a target s gist trace given a related probe; V t probability of retrieving a target s verbatim trace given a target probe; V r probability of retrieving a target s verbatim trace given a related probe. the modified CR model responded to the manipulations as predicted, supporting the conclusion that they provide valid indicators of gist memory. Experiments 3 and 4: Verbatim Memory Next, we turned to validating the verbatim memory parameters, using the same manipulation that has been used to validate verbatim parameters in the original CR paradigm. In Experiments 3 and 4, half of the targets were presented repeatedly on the study list. This should increase verbatim memory for these targets, as an additional presentation doubles the time of exposition of the perceptual surface of that stimulus. Repetition has been shown to selectively affect verbatim memory in the original CR paradigm (Brainerd et al., 1999, Experiments 1 and 2), and we aimed to replicate this finding within the simplified CR paradigm. We predicted that the V t parameter would respond to the manipulation of verbatim trace strength by repeated presentation. Recall that this parameter is an indicator for the identity process by which targets identity is verified with the help of a successfully retrieved verbatim trace. Gist memory should be affected by this manipulation only to a negligible extent because a single presentation of a word usually suffices to fully extract its meaning and activate a gist representation, and an additional presentation would not add to this activation. Method Participants. Twenty volunteers participated in each experiment (Experiment 3: 15 women; ages ranged from 18 to 26 years, M 21; Experiment 4: 11 women; ages ranged from 19 to 45 years, M 25). Design. A 2 (Number of Presentations: 1 vs. 2) 3 (Probe Type: target, related, unrelated) repeated-measures design was implemented. Materials. Experiment 3 used the synonym pairs described in the General Method section. The study list presented one randomly selected item from each of the 20 pairs. These items were randomly split into 10 items that were presented once and 10 items that were presented twice. In total, 30 items were presented in random order. At test, the 10 single-presentation items and the 10 repeatedpresentation items were randomly split into five items each that were shown as target probe and five items for which their synonym was presented as related probe. Ten additional occupation names not presented at study were presented as unrelated probes. Order of presentation of these 30 items was randomized. All randomizations were carried out for each participant anew. Experiment 4 used the categories described in the General Method section. The study list presented exemplars from 20 randomly selected categories that were randomly split into two halves of 10 categories each. Categories from the first half were represented by a single exemplar that was presented once; categories from the second half were represented by a single exemplar that was presented twice on the study list. In total, 30 items were presented in random order. At test, the 10 single-presentation categories and the 10 repeated-presentation categories were randomly split into five categories each for which the presented exemplar was shown as target

8 A SIMPLIFIED CONJOINT RECOGNITION PARADIGM 577 probe and five categories for which the nonpresented exemplar was shown as related probe. Ten unrelated probes were randomly selected from the exemplars of the remaining categories that were not presented at study, one exemplar representing each such category. Order of presentation of these 30 items was randomized. All randomizations were carried out for each participant anew. Procedure. The procedures were identical to those of Experiments 1 and 2, with the exception that the memory test followed immediately after the arithmetic filler task. Results Parameter estimates and significance tests are given in Table 3. The predicted effect of repetition was observed on verbatim memory for targets, V t. In both experiments, V t was larger for twicepresented targets than for once-presented targets. No other effects were obtained. As predicted, gist parameters were not affected by the presentation manipulation (all ps.18). Recollection rejection (parameter V r ) was also not affected (both ps.72). Table 3 Estimates (and 95% Confidence Intervals) for the Parameters of the Simplified Conjoint Recognition Model for Experiments 3 and 4 Parameter Target presentations 1 2 Experiment 3 a.16 (.04,.28) b.17 (.12,.22) G t.51 (.25,.77).74 (.27, 1.00) G r (.64,.98) (.60,.98) V t (.69,.87) V r.00 (.00,.76) (.91, 1.00).09 (.00,.79) Experiment 4 a.17 (.06,.29) b.20 (.14,.26) G t.69 (.42,.96).63 (.01, 1.00) G r (.00,.59) (.17,.85) V t (.76,.92) V r.40 (.13,.67) (.93, 1.00).44 (.08,.79) G 2 (df 1) p Note. a probability of guessing target ; b probability of guessing that an item is either a target or a related probe; G t probability of retrieving a target s gist trace given a target probe; G r probability of retrieving a target s gist trace given a related probe; V t probability of retrieving a target s verbatim trace given a target probe; V r probability of retrieving a target s verbatim trace given a related probe. Discussion In Experiments 3 and 4, verbatim memory for target probes was manipulated. During the presentation phase, half of the items were presented once, whereas the other half were presented twice. As predicted, the repetition manipulation affected targets verbatim memory (V t ) but left gist memory parameters unaffected. This replicates the finding by Brainerd et al. (1999) that verbatim memory is affected by repeated presentation in the original CR paradigm. It is concluded that the simplified CR paradigm and model is well capable of assessing targets verbatim trace strength. Recollection rejection (V r ) tended to be increased in Experiment 4 as compared with Experiment 3. This nonsignificant tendency cannot be attributed to differences in verbatim memory strength, as V t was at comparable levels in both experiments. Instead, it might reflect differences in metacognitive strategy use in the recollection rejection decision (e.g., Gallo, 2004). Recollection rejection also tended to be higher for repeatedly presented items but not significantly so. This is in line with previous research in which the probability of recollection rejection has been increased by repeated presentation of targets in some experiments but to a smaller extent than verbatim memory for targets (e.g., Brainerd et al., 1999; Lampinen et al., 2004). To demonstrate that the simplified CR paradigm is capable of measuring recollection rejection, we conducted two additional experiments using the priming manipulation introduced by Brainerd et al. (1999, Experiment 3). Experiments 5 and 6: Recollection Rejection In Experiments 5 and 6, the process of recollection rejection by way of a nonidentity judgment was examined. Recollection rejection occurs when participants are confronted with a related probe and succeed in retrieving verbatim information for the corresponding target. This verbatim information can then be compared with the probe and will result in a judgment of nonidentity and a correct rejection of the probe (or, in case of the present paradigm, in a correct classification as a related probe). To manipulate the probability of occurrence of this process, we replicated the priming manipulation that was used by Brainerd et al. (1999, Experiment 3) to demonstrate the recollection rejection process. In the memory test, three conditions were realized: One third of participants were presented with the target probe just before the corresponding related probe was tested (target-first condition). This manipulation primes the verbatim trace of the target and should render a recollection rejection due to a nonidentity judgment more likely as compared with a second condition in which target items were tested after the corresponding related probe (target-last condition). A Bayesian Framework for Guessing in the Simplified CR Paradigm Experiments 5 and 6 also aimed at validating the guessing parameter a. For this purpose, a third condition was realized in which targets corresponding to related probes were not presented at all (control condition), and the base rate of targets was therefore

9 578 STAHL AND KLAUER only half that of the target-first condition. In this condition, estimates of parameter a were predicted to be smaller than in the target-first condition. For the target-last condition, intermediate values of a were predicted. These predictions were derived from a Bayesian framework of metacognitive guessing strategies (Batchelder & Batchelder, 2008) as explained in the following paragraphs. In the simplified CR paradigm, two mnemonic states can be distinguished in which the target/related guessing process modeled by parameter a affects performance. Let m 1 be the state in which gist but no verbatim memory is available. Further, let m 2 be the state in which neither verbatim nor gist memory are available, but a given probe s gist has been guessed old as described by parameter b. In both states, participants have decided against the unrelated response but have to decide between the target and related response options. In the absence of relevant information in memory, this decision is guided by a guessing process modeled by parameter a. In a given mnemonic state m i, participants can determine the optimal response category k by computing the conditional probability p(k m i ), for the two response categories target and related. The optimal response is to select the response category k for which p(k m i ) is maximal. In states m 1 and m 2, a guessing-based decision between the response options target and related has to be made. This decision is modeled by parameter a in both states, and therefore, a joint strategy is determined for guessing in m 1 and m 2 to predict values of parameter a. Such a strategy would advise a target response when p(target m 1 m 2 ) is larger than p(related m 1 m 2 ), or stated differently, when the Bayes factor BF a p(target m 1 m 2 )/p(related m 1 m 2 ) is greater than 1. From p(target m 1 m 2 ) p(m 1 m 2 target)p(target)/p(m 1 m 2 ) and p(related m 1 m 2 ) p(m 1 m 2 related)p(related)/p(m 1 m 2 ), and the model equations, it follows that BF a ( p (t) /p (r) )[(1 V t )/(1 V r )][G t (1 G t )b]/[g r (1 G r )b], with p (t) proportion of targets in the test list, and p (r) proportion of related distracters in the test list. As indicated by the first term, the tendency to guess target is expected to increase with the proportion of target probes in the test list, and it is expected to decrease with an increasing proportion of related probes. This reflects a guessing strategy based on base rates. Note that in the case of zero memory, the Bayes factor simplifies to a simple base rate ratio, BF a p (t) /p (r). The second term indicates that the tendency to guess target is expected to increase when verbatim memory for target items (V t ) decreases and when the probability for recollection rejection (V r ) increases. As verbatim memory for target items increases (i.e., V t V r ), optimal guessing shifts toward related probes. Analogously, if V r V t, then optimal guessing shifts toward the target response category. Thus, the ratio of verbatim memory parameters affects the optimal guessing strategy such that guessing is expected to be biased toward the class of items for which verbatim memory is weakest. As indicated by the third term, gist memory exerts an opposite but somewhat weaker effect on the optimal guessing strategy. As the asymmetry between gist memory parameters for the two classes of items increases, so should the guessing tendency toward the class of items with stronger gist memory. The magnitude of this effect is comparable with that of the verbatim memory asymmetry only for b 0, but it is rapidly attenuated as values of b depart from zero. The first two terms are relevant for the conditions tested in Experiments 5 and 6, in which verbatim memory parameters and targets base rates were varied but not gist memory. The second term indicates that the probability of guessing target is expected to be higher in conditions with higher values of V r (i.e., the target-first conditions) than in conditions with low values of V r (i.e., the target-last and control conditions). The first term indicates that the probability of guessing target is expected to be higher in conditions with higher proportions of target probes (i.e., the target-first and target-last conditions) than in conditions with equal proportions of target and related probes (i.e., the control condition). Participants guessing tendency is expected to be influenced by a combination of both effects. Taken together, the magnitude of estimates of parameter a is expected to follow the order: target-first target-last control. Method Participants. Sixty volunteers participated in each experiment (Experiment 5: 42 women; ages ranged from 18 to 49 years, M 24; Experiment 6: 32 women; ages ranged from 19 to 30 years, M 23). Design. A 3 (Priming: target-first, target-last, control) 3 (Probe Type: target, related, unrelated) design was implemented with repeated measures on the last factor. Materials. In Experiment 5, we used the synonym pairs; in Experiment 6, we used the categories described in the General Method section. In contrast to previous experiments, only one item was presented on the study list for each synonym pair and category, and items were presented only once. Order of presentation of the study list was randomized; order of presentation was also randomized for the test list but with the restrictions described below. Randomizations were carried out for each participant anew. In Experiment 5, 20 synonym pairs were represented on the study list by a single item that was randomly selected. At test, these pairs were randomly split into two halves of 10 pairs each. For the first half, the item presented on the study list was shown as target probe. For the second half, the nonpresented item of the pair was shown as related probe and the presented item was used to implement the priming manipulation in the target-first and targetlast conditions. In Experiment 6, 20 categories were randomly selected to be represented on the study list by a single, randomly selected exemplar. At test, these pairs were randomly split into 10 categories for which the presented exemplar was shown as target probe and 10 categories for which the nonpresented exemplar was shown as related probe and the presented exemplar was used to implement the priming manipulation. Procedure. Procedure was identical to that in Experiments 3 and 4 in the control condition. Departing from previous procedure, a priming manipulation was introduced in the memory test for the target-first and target-last conditions of Experiments 5 and 6. The priming manipulation was implemented as follows: In the targetfirst condition, before a given related probe was shown on the test list, the target to which it is related was probed. In contrast, in the target-last condition, the related probe was shown on the test list before the corresponding target was probed. In Experiment 5,